Abstract:

This study tests whether machine learning can forecast PM2.5 air pollution in Almaty 24 hours in advance. Three baseline methods and three ML models (Ridge, Random Forest, XGBoost) were compared on seven months of hourly sensor data. The simple 24-hour moving average outperformed every ML model — a seasonal shift between winter training data and summer test data meant the more complex models systematically overestimated. The practical takeaway: model complexity does not guarantee better forecasts.